Initial Investigation of Kolmogorov–Arnold Networks and Fractional Deep Learning for Personalized Prosthetic Control: A Pilot Study

This article investigates the potential of learnable activation functions in Kolmogorov–Arnold Networks (KANs) for personalized control in a lower-limb prosthesis. Subject-specific versus pooled training data is evaluated to improve machine learning (ML) and deep learning (DL) prediction of turn intent. Inertial measurement unit (IMU) data were collected from five individuals with lower-limb amputation performing turning tasks in a laboratory setting while using the Variable Stiffness Torsion Adapter. The ability to classify an upcoming turn was evaluated using a Multilayer Perceptron (MLP) and a KAN, as well as deep learning models including Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), Convolutional Neural Network–Long Short-Term Memory (CNNLSTM), Gated Recurrent Units (GRU), and their corresponding fractional variants (FCNN, FLSTM, FCNNLSTM, FGRU). Models were trained separately on subject-specific and pooled data to evaluate the impact of training data on performance. Models were deployed on a Raspberry Pi 3 to evaluate real-time inference speed and computational feasibility. On average, the proposed DL models with learnable activation functions increased macro-F1 by 3.54%. While learnable activation functions provided the highest functional improvement in the FCNNLSTM architecture, formal statistical significance was limited by strict multiple comparison corrections. Conventional ML model (MLP) required subject-specific training for optimal performance. Conversely, DL models trained on pooled data achieved functional parity with those trained on subject-specific data. A reduction in clinical burden can be achieved by employing DL techniques on pooled data, which eliminates the need for subject-specific calibration. All models met the real-time performance requirements when deployed on the Raspberry Pi 3.